Classic applications become the deterministic core that agents and people work on: API and MCP first, UI on demand, replacing data entry with finding. With my receipts agent and HourPay as examples.
How far does an AI agent get in a real setup, and what has to be in place first? A gateway, policies, external secrets and two mistakes that could have hurt.
I let an autonomous AI agent build a small, genuinely useful tool from start to finish. Three tasks, three runs, and one clear lesson about where the loop carries the work and where a human still has to set the direction.
Half the industry says: stop writing prompts, write loops. After two weeks on my own code I know better: the loop is the cheap part. The real work is the verifiers. And why coding is the easy case.
For a few weeks now something has been shifting in my daily work with coding agents. A few honest thoughts on Loop Engineering – about token costs, controllability, and why craftsmanship is becoming more important right now, not less.
AI has become a premise in many places rather than the result of deliberation. On vibe coding, sprawl in large organisations, and the question that is rarely asked: do we actually need this?
The EU AI Act does not only affect tech giants. What small and medium enterprises in Austria, Germany, and Switzerland actually need to do, and what they can skip.
Specifications aren't bureaucracy. They are the most effective tool for building software that does what it's supposed to, especially in the age of AI-generated code.
Programming is code production. Development is systems thinking plus responsibility. In the age of AI, this distinction matters more than ever.